AnHai DoanUniversity of Wisconsin

Big Data, Big Knowledge, and Big Crowd

2

The world has changed; now everything is data centric– everyone collects, stores, analyzes TBs and PBs of data

To manage data in this new world, need 3B technologies: – lot of data need big data technologies to scale up algorithms

– data is noisy, unstructured, heterogeneous need a lot of domain knowledge to understand such knowledge is often captured in big knowledge bases

– algorithms are imperfect, certain things humans do better, need humans in the loop, scale is such that there is not enough human developers need crowdsourcing with big crowd

3

Examples

Semantic analysis of the Twitter stream– process 3000-6000 tweets per sec, need fast data infrastructure– to recognize entities, e.g., “go giant!”, need a big KB– KB being built in real time using crowdsourcing

Product matching for e-commerce– build 500+ matchers to match products

one matcher per category: toy, electronics, clothes, etc.– match 500K electronics products with 500K need Hadoop– use a KB to match numerous synonyms: soft cover = paperback,

etc.– use crowdsourcing to generate training and testing data

4

Big Knowledge Technologies Everyone is now building KBs

– IT companies: Google, Microsoft, …– e-retailers: Amazon, Walmart, … – stodgy behemoths: Johnson Control, GE, … – tiny startups, academia, …

User communities are building KBs (e.g., biomedical) There will be not just data centers, but also knowledge

centers– KBs and tools that use such KBs– critical for understanding data (e.g., tweets)

How do we help people build KBs? Knowledge centers? – a next important direction for data integration research

5

Big Crowd Technologies Industry has been doing these for years For us it’s not a fad, it’s fundamental

– as data management increasingly involves semantic problems

Have gotten off to a good start (platforms / problems) Need hands-off crowdsourcing

– no developer in the loop, otherwise will not scale– e.g., crowdsourcing 500 product matching problems, one per category

Need crowdsourcing for the masses– e.g., journalist wants to match two political lists of donors

Need “grand challenges” for crowdsourcing? – e.g., something like Wikipedia?